1.AXS_baseline.py
# %%
import warnings
warnings.filterwarnings("ignore")
from airbot.backend import Arm, Camera, Base, Gripper
import os
import numpy as np
import copy
from airbot.backend.utils.utils import camera2base, armbase2world
from airbot.lm import Detector, Segmentor
from airbot.grasp.graspmodel import GraspPredictor
from PIL import Image
import time
import cv2
from airbot.example.utils.draw import draw_bbox, obb2poly
from airbot.example.utils.vis_depth import vis_image_and_depth
from scipy.spatial.transform import Rotation
from threading import Thread, Lock
import math
from airbot.lm.utils import depth2cloud
os.environ['LM_CONFIG'] = "/root/Workspace/AXS_baseline/ICRA2024-Sim2Real-AXS/local.yaml"
os.environ['CKPT_DIR'] = '/root/Workspace/AXS_baseline/ckpt'
class Solution:
CAMERA_SHIFT = np.array([-0.093, 0, 0.07]), np.array([0.5, -0.5, 0.5, -0.5])
BEFORE_MW_BASE_POSE = (np.array([
-0.02596185755302699,
-0.5065354084074988,
-0.17365354084074988,
]), np.array([
0.00179501,
0.06667202,
0.04863613,
0.99658725,
]))
ARM_POSE_TO_MICROWAVE = (np.array([0.67335, 0.2, -0.07678]), np.array([0.0, 0.0, 0.0, 1.0]))
POSE_CLOSE_MICROWAVE = (np.array([-0.05931371154832413, 0.10621326743872146, 0.052710225767309826]), np.array([0.00476639, 0.08115882, 0.05789078, 0.99500713]))
ARM_POSE_CLOSE_MICROWAVE = (np.array([0.47335, 0.1, 0.05]), np.array([0.0, 0.0, 0.0, 1.0]))
ARM_POSE_CLOSE_MICROWAVE_END = (np.array([0.4, -0.55, 0.05]), np.array([0.0, 0.0, 0.0, 1.0]))
POSE_TO_BOWL = (np.array([-0.01226429675289562, 0.11259263609492792, -0.042619463119529605]),
np.array([-0.05418989389554988, 0.056031992518506414, 0.707, 0.707]))
ARM_POSE_TO_LOWER_CABINET = (np.array([0.62335, 0.25, -0.05678]), np.array([0.0, 0.0, 0.0, 1.0]))
ARM_POSE_PUT_LOWER = (np.array([0.62335, 0.25, -0.09678]), np.array([0.0, 0.0, 0.0, 1.0]))
ARM_POSE_TO_UPPER_CABINET = (np.array([0.62335, 0.25, 0.22]), np.array([0.0, 0.0, 0.0, 1.0]))
ARM_POSE_PUT_UPPER = (np.array([0.62335, 0.25, 0.17]), np.array([0.0, 0.0, 0.0, 1.0]))
POSE_OPEN_CAB = (np.array([-0.2741637241544763, 0.3241416117180577, -0.07743623649227918]), np.array([-0.026456952502619244, 0.022510511678367467, 0.6642090190392874, 0.7467393692280592]))
ARM_POSE_STANDARD_MOVING = (np.array([0.3225, 0.00, 0.219]), np.array([0.0, 0.0, 0.0, 1.0]))
GRASP_POSE_1 = (np.array([0.345502073568463, 0.49365995419306763, 0.07947950001408821]), np.array(
[-0.0051582, 0.09461529, 0.05903872, 0.99374834]))
GRASP_POSE_2 = (np.array([0.40524870062559315, 0.4479303912730242, 0.004671858854359251]), np.array(
[-0.04836788, 0.0417043, 0.66597635, 0.74323402]))
OBSERVE_ARM_POSE_1 = (np.array([
0.2565699,
0.2,
0.171663168,
]), np.array([
-0.13970062182177911,
0.6487791800204252,
0.032918235938941776,
0.7473190092439113,
]))
OBSERVE_ARM_POSE_2 = (np.array([
0.2565699,
-0.2,
0.171663168,
]), np.array([
-0.13970062182177911,
0.6487791800204252,
0.032918235938941776,
0.7473190092439113,
]))
END_POSE = (np.array([-1.2317611908186263, -0.9177336410440479, -0.2374505629662929]),
np.array([0.006592923324957343, -0.012942241749323641, 0.014944697147015459, 0.9997828203003477]))
def __init__(self):
self.arm = Arm(backend='ros')
self.base = Base(backend='ros')
self.gripper = Gripper(backend='ros')
self.gripper.open()
self.camera = Camera(backend='ros')
self.detector = Detector(model='grounding-dino')
# self.detector = Detector(model='yolo-v7')
self.segmentor = Segmentor(model='segment-anything')
self.grasper = GraspPredictor(model='graspnet')
self.image_lock = Lock()
self.result_lock = Lock()
self.prompt_lock = Lock()
self.running = True
self.prompt = 'sky'
self.update_once()
self.t_vis = Thread(target=self.vis, daemon=True)
self.t_vis.start()
self.t_update = Thread(target=self.update, daemon=True)
self.t_update.start()
@property
def image(self):
with self.image_lock:
return copy.deepcopy(self._image)
@image.setter
def image(self, value):
with self.image_lock:
self._image = copy.deepcopy(value)
@property
def prompt(self):
with self.prompt_lock:
return copy.deepcopy(self._prompt)
@prompt.setter
def prompt(self, value):
with self.prompt_lock:
self._prompt = copy.deepcopy(value)
@property
def depth(self):
with self.image_lock:
return copy.deepcopy(self._depth)
@depth.setter
def depth(self, value):
with self.image_lock:
self._depth = copy.deepcopy(value)
@property
def bbox(self):
with self.result_lock:
return copy.deepcopy(self._bbox)
@bbox.setter
def bbox(self, value):
with self.result_lock:
self._bbox = copy.deepcopy(value)
@property
def mask(self):
with self.result_lock:
return copy.deepcopy(self._mask)
@mask.setter
def mask(self, value):
with self.result_lock:
self._mask = copy.deepcopy(value)
def update_once(self):
with self.image_lock, self.result_lock:
self._image = copy.deepcopy(self.camera.get_rgb())
self._depth = copy.deepcopy(self.camera.get_depth())
self._det_result = self.detector.infer(self._image, self._prompt)
self._bbox = self._det_result['bbox'].numpy().astype(int)
self._sam_result = self.segmentor.infer(self._image, self._bbox[None, :2][:, [1, 0]])
self._mask = self._sam_result['mask']
def update(self):
while self.running:
self.update_once()
time.sleep(0.005)
def vis(self):
try:
# Infinite loop to display images
while self.running:
image_draw = self.image
image_draw = image_draw * (self.mask[:, :, None].astype(np.uint8) * 0.75 + 0.25)
image_draw = draw_bbox(image_draw, obb2poly(self.bbox[None, ...]).astype(int))
image_draw = image_draw.astype(np.uint8)
image_show = cv2.cvtColor(image_draw, cv2.COLOR_RGB2BGR)
cv2.imshow('RGB', image_show)
_flag = np.any(self.mask)
if _flag == True:
cv2.putText(image_show,
f"det score: {self._det_result['score']}, sam score: {self._sam_result['score']}",
(10, 50), cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 255, 255), 2)
cv2.imshow('RGB', image_show)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
except KeyboardInterrupt:
print("Exiting due to user interruption.")
finally:
cv2.destroyAllWindows()
@staticmethod
def _bbox2mask(image, bbox):
mask = np.zeros_like(image[:, :, 0], dtype=bool)
mask[
bbox[0] - bbox[2] // 2:bbox[0] + bbox[2] // 2,
bbox[1] - bbox[3] // 2:bbox[1] + bbox[3] // 2,
] = True
return mask
@staticmethod
def base_cloud(image, depth, intrinsic, shift, end_pose):
cam_cloud = depth2cloud(depth, intrinsic)
cam_cloud = np.copy(np.concatenate((cam_cloud, image), axis=2))
return camera2base(cam_cloud, shift, end_pose)
def grasp(self):
# find the center of self.mask
# center_cam = np.array(np.nonzeros(self.mask)).mean(axis=1)
# center_frame = self.base_cloud[center_cam]
# self.base.move_to(*self.OPERATE_BASE_POSE_1)
# self.arm.move_end_to_pose(*self.OPERATE_ARM_POSE)
# time.sleep(3)
# method = input('method')
method = "2"
with self.image_lock, self.result_lock:
_depth = copy.deepcopy(self._depth)
_image = copy.deepcopy(self._image)
_bbox = copy.deepcopy(self._bbox)
_mask = copy.deepcopy(self._mask)
cloud = self.base_cloud(_image, _depth, self.camera.INTRINSIC, self.CAMERA_SHIFT, self.arm.end_pose)
if method == "1":
direction = cloud[_bbox[0] - _bbox[2] // 2, _bbox[1]][:3] - self.arm.end_pose[0]
direction = direction / np.linalg.norm(direction)
grasp_position = cloud[_bbox[0] - _bbox[2] // 2 + 9, _bbox[1]][:3] - 0.12 * direction
grasp_rotation = Rotation.from_euler('xyz', [0, np.pi / 2, np.pi / 2], degrees=False).as_quat()
elif method == "2":
grasp_position = cloud[ _bbox[0], _bbox[1] - _bbox[3] // 2 + 8][:3]
grasp_position[2] = -0.165
grasp_rotation = Rotation.from_euler('xyz', [0, np.pi / 2, 0], degrees=False).as_quat()
else:
bbox_mask = self._bbox2mask(_image, _bbox)
(grasp_position, grasp_rotation), _ = self.grasper.infer(cloud, bbox_mask)
grasp_rotation = Rotation.from_euler("yz", [np.pi / 2, np.pi / 2], degrees=False).as_quat()
self.arm.move_end_to_pose(grasp_position, grasp_rotation)
time.sleep(2)
self.gripper.close()
time.sleep(4)
self.arm.move_end_to_pose(*self.ARM_POSE_STANDARD_MOVING)
@staticmethod
def _vis_grasp(cloud, position, rotation):
import open3d as o3d
from graspnetAPI.grasp import GraspGroup
o3d_cloud = o3d.geometry.PointCloud()
cloud = copy.deepcopy(cloud)
o3d_cloud.points = o3d.utility.Vector3dVector(cloud[:, :, :3].reshape(-1, 3).astype(np.float32))
o3d_cloud.colors = o3d.utility.Vector3dVector(cloud[:, :, 3:].reshape(-1, 3).astype(np.float32) / 255.)
gg = GraspGroup(
np.array([1., 0.06, 0.01, 0.06, *Rotation.from_quat(rotation).as_matrix().flatten(), *position,
0]).reshape(1, -1))
coordinate_frame = o3d.geometry.TriangleMesh.create_coordinate_frame(size=0.5, origin=[0, 0, 0])
o3d.visualization.draw_geometries([o3d_cloud, *gg.to_open3d_geometry_list(), coordinate_frame])
def place_microwave(self):
self.base.move_to(*self.BEFORE_MW_BASE_POSE, 'world', False)
self.arm.move_end_to_pose(*self.ARM_POSE_TO_MICROWAVE)
input_pose = (np.array([0.25, 0.0, 0.0]), np.array([0.0, 0.0, 0.0, 1.0]))
self.base.move_to(*input_pose, 'robot', True)
self.gripper.open()
time.sleep(2)
output_pose = (np.array([-0.3, 0.0, 0.0]), np.array([0.0, 0.0, 0.0, 1.0]))
self.base.move_to(*output_pose, 'robot', True)
def close_microwave(self):
self.arm.move_end_to_pose(*self.ARM_POSE_STANDARD_MOVING)
self.base.move_to(*self.POSE_CLOSE_MICROWAVE, 'world', False)
self.arm.move_end_to_pose(*self.ARM_POSE_CLOSE_MICROWAVE)
self.arm.move_end_to_pose(*self.ARM_POSE_CLOSE_MICROWAVE_END)
output_pose1 = (np.array([-0.25, 0.0, 0.0]), np.array([0.0, 0.0, 0.0, 1.0]))
self.base.move_to(*output_pose1, 'robot', True)
self.arm.move_end_to_pose(*self.ARM_POSE_STANDARD_MOVING)
def place_bowl_lower(self):
self.arm.move_end_to_pose(*self.ARM_POSE_STANDARD_MOVING)
self.base.move_to(*self.POSE_TO_BOWL, 'world', False)
self.arm.move_end_to_pose(*self.ARM_POSE_TO_LOWER_CABINET)
input_pose = (np.array([0.35, 0.0, 0.0]), np.array([0.0, 0.0, 0.0, 1.0]))
self.base.move_to(*input_pose, 'robot', True)
self.gripper.open()
self.arm.move_end_to_pose(*self.ARM_POSE_PUT_LOWER)
time.sleep(2)
output_pose = (np.array([- 0.35, 0.0, 0.0]), np.array([0.0, 0.0, 0.0, 1.0]))
self.base.move_to(*output_pose, 'robot', True)
self.arm.move_end_to_pose(*self.ARM_POSE_STANDARD_MOVING)
def place_bowl_upper(self):
self.arm.move_end_to_pose(*self.ARM_POSE_STANDARD_MOVING)
self.base.move_to(*self.POSE_TO_BOWL,'world', False)
self.arm.move_end_to_pose(*self.ARM_POSE_TO_UPPER_CABINET)
input_pose = (np.array([0.35, 0.0, 0.0]), np.array([0.0, 0.0, 0.0, 1.0]))
self.base.move_to(*input_pose, 'robot', True)
self.gripper.open()
self.arm.move_end_to_pose(*self.ARM_POSE_PUT_UPPER)
time.sleep(2)
output_pose = (np.array([-0.35, 0.0, 0.0]), np.array([0.0, 0.0, 0.0, 1.0]))
self.base.move_to(*output_pose, 'robot', True)
self.arm.move_end_to_pose(*self.ARM_POSE_STANDARD_MOVING)
def lookforonce(self, det_th, sam_th):
with self.image_lock, self.result_lock:
_rgb = copy.deepcopy(self.camera.get_rgb())
_depth = copy.deepcopy(self.camera.get_depth())
_det_result = copy.deepcopy(self._det_result)
_sam_result = copy.deepcopy(self._sam_result)
_bbox = _det_result['bbox'].numpy().astype(int)
_mask = _sam_result['mask']
if np.any(_mask) is False:
print(f"direction {direction} not Found")
print("det score:", _det_result['score'])
print("sam score:", _sam_result['score'])
if _det_result['score'] > det_th and _sam_result['score'] > sam_th:
print(f"Found the {self._prompt}")
centerpoint = depth2cloud(_depth, self.camera.INTRINSIC, organized=True)[_bbox[0] // 1, _bbox[1] // 1]
centerpoint = camera2base(centerpoint, self.CAMERA_SHIFT, self.arm.end_pose)
centerpoint = (armbase2world(centerpoint, (self.base.position, self.base.rotation)).squeeze())
object_rgb = _rgb[_bbox[0] - np.int32(_bbox[2]/4):_bbox[0] + np.int32(_bbox[2]/4), _bbox[1] - np.int32(_bbox[3]/4):_bbox[1] + np.int32(_bbox[3]/4)]
mean_rgb = (np.mean(np.mean(object_rgb, axis=0), axis=0).astype(int))
print('-' * 50)
print('centerpoint is', centerpoint)
print('object rgb is', mean_rgb)
print('-' * 50)
return centerpoint, mean_rgb
if __name__ == '__main__':
s = Solution()
s.base.move_to(*s.POSE_OPEN_CAB, 'world', False)
time.sleep(1)
POS_DOOR_HANDLE = np.array([0.30353946626186371, 1.230472918510437, 0])
centerp_car = np.linalg.inv(np.array(Rotation.from_quat(s.base.rotation).as_matrix())).dot((POS_DOOR_HANDLE-s.base.position))
ARM_POSE_DOOR_HANDLE = (np.array([
centerp_car[0] - 0.2975 - 0.01,
centerp_car[1] + 0.17309 - 0.01,
0.2,
]), np.array([
0.0,
0.0,
0.0,
1.0,
]))
s.arm.move_end_to_pose(*ARM_POSE_DOOR_HANDLE)
s.gripper.close()
time.sleep(5)
for i in range(7):
d = 0.1 * (i + 1)
new_ori = np.array([0, 0, 0, 1])
new_pos = ARM_POSE_DOOR_HANDLE[0] + np.array([-0.4*math.sin(d), 0.4-0.4*math.cos(d), 0])
r = Rotation.from_euler("xyz", np.array([0, 0, -d]), degrees=False)
new_ori = r.as_quat()
s.arm.move_end_to_pose(new_pos, np.array(new_ori))
time.sleep(0.5)
s.gripper.open()
time.sleep(3)
s.arm.move_end_to_pose(np.array([0.3225, 0.00, 0.219]), np.array([0.0, 0.0, 0.0, 1.0]))
s.arm.move_end_to_pose(np.array([0.3225, -0.25, 0.219]), np.array([0.0, 0.0, 0.0, 1.0]))
s.arm.move_end_to_pose(np.array([0.5615004168820418, -0.2, 0.35123932220414126]), np.array([0.0, 0.0, 0.2953746452532359, 0.9547541169761965]))
s.arm.move_end_to_pose(np.array([0.6015004168820418, -0.15, 0.35123932220414126]), np.array([0.0, 0.0, 0.2953746452532359, 0.9547541169761965]))
s.arm.move_end_to_pose(np.array([0.4882092425581316, 0.2917225555849343, 0.3515424067641672]), np.array([0.0, 0.0, 0.6045684271573144, 0.7957869908463996]))
back_pose = (np.array([-0.05, 0.0, 0.0]), np.array([0.0, 0.0, 0.0, 1.0]))
s.base.move_to(*back_pose, 'robot', True)
s.arm.move_end_to_pose(*s.ARM_POSE_STANDARD_MOVING)
s._prompt = 'white mug'
cp = None
s.base.move_to(*s.GRASP_POSE_1, 'world', False)
look_num = 0
while cp is None:
for direction in [1, 2]:
if direction == 1:
s.arm.move_end_to_pose(*s.OBSERVE_ARM_POSE_1)
else:
s.arm.move_end_to_pose(*s.OBSERVE_ARM_POSE_2)
cp = s.lookforonce(0.65,0.65)
if cp is not None:
break
look_num += 1
if look_num>3:
break
if cp is not None:
centerpoint, object_mean_rgb = cp
centerp_car = np.linalg.inv(np.array(Rotation.from_quat(s.base.rotation).as_matrix())).dot((centerpoint-s.base.position))
OBSERVE_ARM_POSE_TOP = (np.array([
centerp_car[0]- 0.2975 - 0.05,
centerp_car[1] + 0.17309,
0.018713334665877806,
]), np.array([
-0.13970062182177911,
0.6487791800204252,
0.032918235938941776,
0.7473190092439113,
]))
s.arm.move_end_to_pose(*OBSERVE_ARM_POSE_TOP)
time.sleep(1)
s.grasp()
s.place_microwave()
s.close_microwave()
obj_rgb = []
for j in range(5):
s._prompt = 'bowl'
cp = None
s.base.move_to(*s.GRASP_POSE_1, 'world', False)
look_num = []
while cp is None:
for direction in [1, 2]:
if direction == 1:
s.arm.move_end_to_pose(*s.OBSERVE_ARM_POSE_1)
else:
s.arm.move_end_to_pose(*s.OBSERVE_ARM_POSE_2)
cp = s.lookforonce(0.6, 0.6)
if cp is not None:
break
look_num.append(1)
if len(look_num)>2:
break
if len(look_num)>2:
break
centerpoint, object_mean_rgb = cp
centerp_car = np.linalg.inv(np.array(Rotation.from_quat(s.base.rotation).as_matrix())).dot((centerpoint-s.base.position))
OBSERVE_ARM_POSE_TOP = (np.array([
centerp_car[0]- 0.2975 - 0.05,
centerp_car[1] + 0.17309,
0.018713334665877806,
]), np.array([
-0.13970062182177911,
0.6487791800204252,
0.032918235938941776,
0.7473190092439113,
]))
s.arm.move_end_to_pose(*OBSERVE_ARM_POSE_TOP)
time.sleep(1)
s.grasp()
obj_rgb.append(object_mean_rgb)
if j != 0:
print("color", abs(sum(obj_rgb[j]-obj_rgb[0])))
if j == 0:
s.place_bowl_lower()
elif abs(sum(obj_rgb[j]-obj_rgb[0]))>30:
s.place_bowl_upper()
else:
s.place_bowl_lower()
obj_rgb_2 = []
for j in range(5):
s._prompt = 'bowl'
cp = None
s.base.move_to(*s.GRASP_POSE_2, 'world', False)
look_num = []
while cp is None:
for direction in [1, 2]:
if direction == 1:
s.arm.move_end_to_pose(*s.OBSERVE_ARM_POSE_1)
else:
s.arm.move_end_to_pose(*s.OBSERVE_ARM_POSE_2)
cp = s.lookforonce(0.6, 0.6)
if cp is not None:
break
look_num.append(1)
if len(look_num)>2:
break
if len(look_num)>2:
break
centerpoint, object_mean_rgb = cp
centerp_car = np.linalg.inv(np.array(Rotation.from_quat(s.base.rotation).as_matrix())).dot((centerpoint-s.base.position))
OBSERVE_ARM_POSE_TOP = (np.array([
centerp_car[0]- 0.2975 - 0.05,
centerp_car[1] + 0.17309,
0.018713334665877806,
]), np.array([
-0.13970062182177911,
0.6487791800204252,
0.032918235938941776,
0.7473190092439113,
]))
s.arm.move_end_to_pose(*OBSERVE_ARM_POSE_TOP)
time.sleep(1)
s.grasp()
obj_rgb_2.append(object_mean_rgb)
if len(obj_rgb) != 0:
if abs(sum(obj_rgb_2[j]-obj_rgb[0]))>30:
s.place_bowl_upper()
else:
s.place_bowl_lower()
else:
if j == 0:
s.place_bowl_lower()
elif abs(sum(obj_rgb_2[j]-obj_rgb_2[0]))>30:
s.place_bowl_upper()
else:
s.place_bowl_lower()
print("finish the task")
s.base.move_to(*s.END_POSE, 'world', False)
time.sleep(20)2.ros_base_control.py
from . import transformations, CoordinateTools
import numpy as np
import time
from typing import Union
from threading import Thread
class BaseControlTools(object):
"""机器人底盘运动控制工具类：
pose(tuple)：
0:position：（x,y,z）
1:rotation：（r,p,y）（欧拉角，按xyz顺序，r:[-pi,pi],p:[-pi/2,pi/2],y:[-pi,pi]）(函数参数设置也可使用四元数，按xyzw顺序)
velocity(tuple)：
0:linear：（x,y,z）
1:angular：（r,p,y）（与rotation对应）
"""
coor = CoordinateTools()
_TEST_ = False
@staticmethod
def target_pose_to_velocity(
posi_or_rota: Union[float, np.ndarray], kp, limits: tuple, dead_zone: float
) -> Union[float, np.ndarray]:
"""根据目标在机器人坐标系下的位置或姿态和给定参数计算机器人的线速度或角速度"""
if limits[0] > limits[1]:
limits = (limits[1], limits[0])
if dead_zone > limits[0]:
dead_zone = limits[0]
if isinstance(posi_or_rota, np.ndarray):
raw_velocity_norm = np.linalg.norm(posi_or_rota)
else:
raw_velocity_norm = abs(posi_or_rota)
target_linear_velocity_norm = kp * raw_velocity_norm
if target_linear_velocity_norm == 0:
target_velocity = posi_or_rota
elif target_linear_velocity_norm > limits[1]:
target_velocity = limits[1] / raw_velocity_norm * posi_or_rota
elif target_linear_velocity_norm < dead_zone:
target_velocity *= 0.0
elif target_linear_velocity_norm < limits[0]:
target_velocity = limits[0] / raw_velocity_norm * posi_or_rota
else:
target_velocity = (
target_linear_velocity_norm / raw_velocity_norm * posi_or_rota
)
return target_velocity
@staticmethod
def min_rotation_move(distance: float, direcition: float) -> tuple:
"""将机器人从当前方向轴按最小方向转向目标方向轴，使得机器人的方向轴（如x轴）与目标方向共线（同向或反向）。
输入：
distance: 机器人与目标点的距离 >= 0
direction: 机器人与目标点的方向 [-pi, pi]
输出：
distance: 机器人与目标点的距离，若转向后机器人方向轴与目标方向同向则为正，反向则为负
direction: 机器人与目标点的方向 [-pi/2, pi/2]
"""
if direcition < -np.pi / 2:
direcition += np.pi
distance *= -1
elif direcition > np.pi / 2:
direcition -= np.pi
distance *= -1
return distance, direcition
@classmethod
def composit_velocity(
cls, target_pose, current_pose, kp, limits, dead_zone
) -> tuple:
"""根据目标位姿和当前位姿及给定参数计算机器人的线速度和角速度"""
raw_linear_velocity, raw_angular_velocity = cls.coor.to_robot_coordinate(
target_pose, current_pose
)
target_linear_velocity = cls.target_pose_to_velocity(
raw_linear_velocity, kp[0], limits[0], dead_zone[0]
)
target_angular_velocity = cls.target_pose_to_velocity(
raw_angular_velocity, kp[1], limits[1], dead_zone[1]
)
return target_linear_velocity, target_angular_velocity
@classmethod
def three_stage_control(
self, target_pose, current_pose, tolerance, kp, limits, dead_zone
) -> tuple:
"""三阶段底盘位置-速度控制法(先旋转调整方向，再平移调整位置，然后再选择调整姿态；根据误差反馈动态调整)"""
target_linear_velocity = np.zeros(3, dtype=np.float64)
target_angular_velocity = np.zeros(3, dtype=np.float64)
stage, bias = self.get_stage_and_bias(
target_pose, current_pose, tolerance[:2], tolerance[2], improve=False
)
if stage in [1, 3]:
target_angular_velocity[2] = self.target_pose_to_velocity(
bias, kp[1], limits[1], dead_zone[1]
)
elif stage == 2:
target_linear_velocity[0] = self.target_pose_to_velocity(
bias, kp[0], limits[0], dead_zone[0]
)
return target_linear_velocity, target_angular_velocity
@classmethod
def get_stage_and_bias(
cls,
target_pose: tuple,
current_pose: tuple,
pose_tolerance: tuple,
direction_tolerance: Union[float, tuple, None] = None,
improve=False,
last_stage=None,
new_target=1,
avoid_321=False,
same_ignore=False,
avoid_swing=False,
) -> tuple:
"""获得三阶段控制法的当前阶段及其相关偏差量"""
position_error, rotation_error = cls.coor.get_pose_error_in_axis(
target_pose, current_pose
) # rotation_error范围为[-pi, pi]
position_distance, rotation_distance = cls.coor.norm(
position_error
), cls.coor.norm(
rotation_error
) # 计算位置以及姿态（优弧范围内）的距离
same_target = same_ignore and new_target == 0
if cls._TEST_:
print("position_error:", position_error)
print("rotation_error:", rotation_error)
print("position_distance:", position_distance)
print("rotation_distance:", rotation_distance)
print("new_target:", new_target)
print("same_target:", same_target)
if (
position_distance <= pose_tolerance[0]
or (avoid_321 and last_stage == 3 and new_target == 0)
or (last_stage == -1 and same_target)
):
if rotation_distance <= pose_tolerance[1]:
return -1, 0 # end and stop
else:
return 3, rotation_error[2]
else:
pose_in_robot = cls.coor.to_robot_coordinate(target_pose, current_pose)
position_in_robot = pose_in_robot[0]
direction_error = cls.coor.get_spherical(position_in_robot)[2]
# 劣弧处理成优弧（position_distance这时带有正负了）
position_distance, direction_error = cls.min_rotation_move(
position_distance, direction_error
)
if cls._TEST_:
print("direction_error:", direction_error)
direction_error_abs = abs(direction_error)
if improve: # 使用改进三阶段控制法
if (
((last_stage in [None, 1, -1]) or not same_target)
and direction_error_abs > direction_tolerance[0]
) or direction_error_abs > direction_tolerance[1]:
return 1, direction_error
elif direction_error_abs <= 0: # 完全对齐后才开始绝对直线
return 2, position_distance
elif direction_error_abs <= direction_tolerance[1]:
if avoid_swing:
direction_error = 0 # 不再修正方向，避免来回摆动
if position_distance < 0:
direction_error *= -1 # 倒车修正
return 1.5, (position_distance, direction_error) # 1.5的bias是个tuple
else:
if (
direction_tolerance is None
and direction_error_abs <= pose_tolerance[2]
) or (direction_error_abs <= direction_tolerance):
return 2, position_distance
else:
return 1, direction_error
@classmethod
def three_stage_control_improved(
self,
target_pose,
current_pose,
pose_tolerance,
direction_tolerance,
kp,
limits,
dead_zone,
enhance=1,
last_stage=None,
new_target=1,
avoid_321=False,
same_ignore=False,
avoid_swing=False,
) -> tuple:
"""改进三阶段底盘位置-速度控制法(增加单轴移动+旋转叠加阶段)"""
target_linear_velocity = np.zeros(3, dtype=np.float64)
target_angular_velocity = np.zeros(3, dtype=np.float64)
stage, bias = self.get_stage_and_bias(
target_pose,
current_pose,
pose_tolerance,
direction_tolerance,
improve=True,
last_stage=last_stage,
new_target=new_target,
avoid_321=avoid_321,
same_ignore=same_ignore,
avoid_swing=avoid_swing,
)
if self._TEST_:
print("stage:", stage)
if stage in [1, 1.5, 3]:
if stage == 1.5:
bias_r = bias[1]
else:
bias_r = bias
target_angular_velocity[2] = (
self.target_pose_to_velocity(bias_r, kp[1], limits[1], dead_zone[1])
/ enhance
)
if stage in [1.5, 2]:
if stage == 1.5:
bias_t = bias[0]
else:
bias_t = bias
target_linear_velocity[0] = (
self.target_pose_to_velocity(bias_t, kp[0], limits[0], dead_zone[0])
* enhance
)
return (target_linear_velocity, target_angular_velocity), stage
@staticmethod
def four_steering_wheel_ik(
linear_vx, linear_vy, angular_vz, width, lenth, radius=1, reduction=1
) -> tuple:
"""四舵轮底盘的逆运动学（给定目标速度和相关参数，计算四轮转向和轮速）"""
r = np.hypot(lenth, width)
# 各轮的合成分速度
A = linear_vy - angular_vz * lenth / r # 右轮的y方向分速度
B = linear_vy + angular_vz * lenth / r # 左轮的y方向分速度
C = linear_vx - angular_vz * width / r # 前轮的x方向分速度
D = linear_vx + angular_vz * width / r # 后轮的x方向分速度
# 各轮合成速度大小
ws1 = np.sqrt(B**2 + C**2)
ws2 = np.sqrt(B**2 + D**2)
ws3 = np.sqrt(A**2 + D**2)
ws4 = np.sqrt(A**2 + C**2)
# 各轮合成速度方向
wa1 = np.arctan2(B, C)
wa2 = np.arctan2(B, D)
wa3 = np.arctan2(A, D)
wa4 = np.arctan2(A, C)
# 车轮自身转速缩放因子（减速比/轮半径）
rotating_factor = reduction / radius
# 各轮转速大小（rotating_factor为1时为线速度）
speeds = np.array([ws1, ws2, ws3, ws4], dtype=np.float64) * rotating_factor
# 各轮旋转角度
angles = np.array([wa1, wa2, wa3, wa4], dtype=np.float64)
return speeds, angles
class BaseControl(object):
"""机器人底盘运动控制顶层类"""
_TEST_ = False
def __init__(self) -> None:
self._move_kp = (100, 100)
self._velocity_dead_zone = (0.0, 0.0)
self._linear_limits = (0.2, 0.3)
self._angular_limits = (0.17, 0.18)
self._direction_tolerance = 0.2
self._direction_tolerance_improved = (0.1, 0.3)
self._wait_tolerance = (0.1, 0.1)
self._improve_enhance = 1
self._last_stage = None
self._wait_timeout = 5
self._wait_period = 0.01
self._move_stop = True
self._move_method = "three_stage_improved"
self._current_position = self._current_rotation = None
BaseControlTools._TEST_ = self._TEST_
self._tools = BaseControlTools()
self._muilti_avoid = {"set": False, "get": False}
self._new_target = 1
self._avoid_321 = False
self._same_ignore = False
self._position_target = np.zeros(3, dtype=np.float64)
self._rotation_target = np.zeros(3, dtype=np.float64)
self._last_orientation_cmd = np.zeros(4, dtype=np.float64)
self._last_pose_ref = "world"
def move_to(self, position: np.ndarray, rotation: np.ndarray) -> bool:
"""设置并移动机器人到指定的目标位姿"""
self.set_target_pose(position, rotation)
result = self.move()
return result
def shift_pose(self, axis: int, target: float) -> bool:
"""在机器人当前位姿的基础上，沿指定轴移动指定距离"""
position, rotation = self.get_current_world_pose()
if axis < 3:
position[axis] += target
else:
rotation[axis-3] += target
result = self.move_to(position, rotation)
return result
def set_target_pose(
self, position: np.ndarray, rotation: np.ndarray, ref: str = "world"
) -> None:
"""
设置机器人在世界（ref=world）/自身（ref=robot）当前坐标系下的目标位姿；
默认为世界坐标系下的目标位姿；
"""
len_rotation = len(rotation)
# 目标相同是否忽略
if self._avoid_321:
if self._last_pose_ref == ref and (position == self._position_target).all():
if len_rotation == 4:
if (rotation == self._last_orientation_cmd).all():
return
elif (rotation == self._rotation_target).all():
return
else:
self._new_target += 1
# 重复设置目标位姿时的警告
if self._muilti_avoid["set"]:
print("Warning: set_target_pose is called before the last is finished!")
return
else:
self._muilti_avoid["set"] = True
# 设置目标位姿
if ref == "robot":
position, rotation = self._tools.coor.to_world_coordinate(
(position, rotation), self.get_current_world_pose()
)
else:
if ref != "world":
print("ref is not 'world' or 'robot', by default 'world' is used!")
ref = "world"
if len_rotation == 4:
self._last_orientation_cmd = rotation
rotation = transformations.euler_from_quaternion(rotation)
rotation = np.array(rotation, dtype=np.float64)
self._last_pose_ref = ref
self._position_target, self._rotation_target = position, rotation
self._muilti_avoid["set"] = False
def get_target_pose(self) -> tuple:
"""获取机器人在世界坐标系下的目标位姿"""
return self._position_target, self._rotation_target
def get_velocity_cmd(self, ignore_stop=False) -> tuple:
"""获取机器人的速度指令"""
if self._muilti_avoid["get"]:
print("Warning: get_velocity_cmd is called before the last is finished!")
return self._vel_cmd
else:
self._muilti_avoid["get"] = True
if self._move_stop and not ignore_stop:
self._vel_cmd = (np.array((0, 0, 0)), np.array((0, 0, 0)))
elif self._move_method == "three_stage":
self._three_stage_control()
elif self._move_method == "three_stage_improved":
self._three_stage_control_improved()
else:
self._composit_velocity()
self._muilti_avoid["get"] = False
if self._new_target > 0:
self._new_target -= 1
return self._vel_cmd
def move(self, time_out=None) -> bool:
"""移动机器人到最新设置的目标位姿"""
self._move_stop = False
start_time = time.time()
self.get_velocity_cmd() # 防止第一次速度指令为0
wait_time = time_out if time_out is not None else self._wait_timeout
while (self._vel_cmd[0] != 0).any() and (self._vel_cmd[1] != 0).any():
if (time.time() - start_time) > wait_time:
print(
"Move timeout, the robot may be stuck! The motion will be stopped and the program will continue!"
)
self._move_stop = True
return False
time.sleep(self._wait_period)
self._move_stop = True
return True
def set_wait_tolerance(
self, position: float, orientation: float, timeout: float, frequency: float
):
"""设置等待的误差容限和超时时间，以及等待的刷新频率"""
self._wait_tolerance = np.array([position, orientation], dtype=np.float64)
self._wait_timeout = timeout
self._wait_period = 1 / frequency
def avoid_321(self, avoid: bool = True, same_ignore: bool = True):
"""避免3阶段旋转误差造成的3阶段中、3阶段结束后相同命令引起的退化"""
self._avoid_321 = avoid
self._same_ignore = same_ignore
def avoid_swing(self, avoid: bool = True):
self._avoid_swing = avoid
def set_move_kp(self, tarns: float, rotat: float):
"""设置机器人运动控制的比例增益"""
self._move_kp = (tarns, rotat)
def set_velocity_limits(self, linear: tuple, angular: tuple):
"""设置机器人运动的线速度和角速度的上下限"""
self._linear_limits = linear
self._angular_limits = angular
def set_velocity_dead_zone(self, linear_dead_zone: float, angular_dead_zone: float):
"""设置机器人运动的线速度和角速度死区"""
self._velocity_dead_zone = (linear_dead_zone, angular_dead_zone)
def set_direction_tolerance(self, tolerance: Union[float, tuple, list]):
"""设置机器人运动的方向容限"""
if isinstance(tolerance, (tuple, list)):
self._direction_tolerance_improved = tolerance
else:
self._direction_tolerance = tolerance
def set_improve_enhance(self, enhance: float):
"""设置机器人运动的方向容限"""
self._improve_enhance = enhance
def set_move_method(self, method: str):
"""设置机器人运动的控制方法"""
self._move_method = method
def _composit_velocity(self):
velocity = self._tools.composit_velocity(
self.get_target_pose(),
self.get_current_world_pose(),
self._move_kp,
(self._linear_limits, self._angular_limits),
self._velocity_dead_zone,
)
self._vel_cmd = velocity
return velocity
def _three_stage_control(self):
velocity = self._tools.three_stage_control(
self.get_target_pose(),
self.get_current_world_pose(),
list(self._wait_tolerance) + [self._direction_tolerance],
self._move_kp,
(self._linear_limits, self._angular_limits),
self._velocity_dead_zone,
)
self._vel_cmd = velocity
return velocity
def _three_stage_control_improved(self) -> tuple:
velocity, self._last_stage = self._tools.three_stage_control_improved(
self.get_target_pose(),
self.get_current_world_pose(),
self._wait_tolerance,
self._direction_tolerance_improved,
self._move_kp,
(self._linear_limits, self._angular_limits),
self._velocity_dead_zone,
enhance=self._improve_enhance,
last_stage=self._last_stage,
new_target=self._new_target,
avoid_321=self._avoid_321,
same_ignore=self._same_ignore,
avoid_swing=self._avoid_swing,
)
if self._TEST_:
print("target_rotation:", self.get_target_pose()[1])
print("current_rotation:", self.get_current_world_pose()[1])
print("target_position:", self.get_target_pose()[0])
print("current_position:", self.get_current_world_pose()[0])
print("velocity cmd:", velocity)
print(" ")
self._vel_cmd = velocity
return velocity
def _get_wait_error(self) -> np.ndarray:
"""得到机器人当前与目标位姿的误差（欧式距离）"""
wait_error = self._tools.coor.get_pose_distance(
self.get_target_pose(), self.get_current_world_pose()
)
return wait_error
def set_control_handel(self, hanle, frequency=None):
"""设置外部控制函数，将启动一个线程来循环执行该函数"""
def thread_handle():
while True:
hanle(self.get_velocity_cmd())
if frequency is not None:
time.sleep(1 / frequency)
Thread(target=thread_handle, daemon=True).start()
def set_current_pose_handle(self, handle):
"""设置获得当前位姿的外部句柄函数调用(重写get_current_world_pose)"""
self.get_current_world_pose = handle
def set_current_world_pose(self, position, rotation):
"""设置机器人当前位姿"""
if len(rotation) == 4:
rotation = transformations.euler_from_quaternion(rotation)
rotation = np.array(rotation, dtype=np.float64)
self._current_position, self._current_rotation = position, rotation
def get_current_world_pose(self) -> tuple:
"""
获取机器人在世界坐标系下的位置和姿态；
若未设置过当前位姿，则返回目标位姿（不论目标位姿是否设置过）；
"""
if self._current_position is None or self._current_rotation is None:
return self._position_target, self._rotation_target
else:
return self._current_position, self._current_rotation
def set_model_four_steering_wheel(self, width, lenth, radius=1, reduction=1):
"""设置四舵轮机器人底盘模型"""
self.chasis_width = width
self.chasis_lenth = lenth
self.wheel_radius = radius
self.wheel_reduction = reduction
def ik_four_steering_wheel(self, linear_vx, linear_vy, angular_vz):
"""四舵轮底盘的逆运动学（给定目标速度和相关参数，计算四轮转向和轮速）"""
speeds, angles = self._tools.four_steering_wheel_ik(
linear_vx,
linear_vy,
angular_vz,
self.chasis_width,
self.chasis_lenth,
self.wheel_radius,
self.wheel_reduction,
)
return speeds, angles
def pose_ik_four_steering_wheel(self):
"""根据当前位姿和目标位姿，直接解算得到四轮转角和转速"""
vel = self._composit_velocity()
speeds, angles = self.ik_four_steering_wheel(vel[0][0], vel[0][1], vel[1][2])
return speeds, angles
if __name__ == "__main__":
import rospy
from geometry_msgs.msg import Pose, Twist
from . import to_ros_msgs
rospy.init_node("test_base_control")
base_control = BaseControl()
base_control.set_move_method("three_stage_improved")
base_control.set_move_kp(1.5, 0.2)
base_control.set_velocity_limits((0.0, 0.3), (0.0, 0.18))
base_control.set_velocity_dead_zone(0.0, 0.0)
base_control.set_direction_tolerance(0.2)
base_control.set_wait_tolerance(0.1, 0.1, 60, 200)
base_control.set_improve_enhance(2)
# 注册机器人底盘外部控制接口
pub = rospy.Publisher("/cmd_vel", Twist, queue_size=10)
def vel_cmd_pub(raw_cmd):
pub.publish(to_ros_msgs.to_Twist(raw_cmd))
base_control.set_control_handel(vel_cmd_pub, frequency=200)
# 实现机器人当前位姿更新接口
def pose_cb(msg: Pose):
euler = transformations.euler_from_quaternion(
[msg.orientation.x, msg.orientation.y, msg.orientation.z, msg.orientation.w]
)
base_control.set_current_world_pose(
np.array(
[msg.position.x, msg.position.y, msg.position.z], dtype=np.float64
),
np.array(euler, dtype=np.float64),
)
rospy.Subscriber("/airbot/pose", Pose, pose_cb)
rospy.sleep(1)
target_pose0 = base_control.get_current_world_pose()
target_pose1 = (target_pose0[0], np.array([0.0, 0.0, 1.5], dtype=np.float64))
target_pose2 = (np.array([-2.5, -4, 0.5], dtype=np.float64), target_pose1[1])
base_control.set_target_pose(*target_pose2)
print("*************************************")
print("Start moving!")
base_control.move()
print("Move finished!")
print("*************************************")
rospy.sleep(0.5)
2061
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
